Raw interaction data, which records what a human does in an educational game, is hard to read. This is as true for large language models as it is for humans. So this project tests converting the raw data to text, which the AI can read. This paper (8 page PDF) reports on a test to see whether r GPT-4o can detect moments of struggle in the data. Plain machine learning "performed better when struggle followed measurable patterns, such as frequent or prolonged pauses between actions," but a combination of ChatGPT and machine learning worked well.
Today: Total: [Direct link] [Share]
